network characteristics of video streaming traffic

1 / 21

Network Characteristics of Video Streaming Traffic

Ashwin Rao†, Yeon-sup Lim*, Chadi Barakat†, Arnaud Legout†, Don Towsley*, and Walid Dabbous†

†INRIA Sophia Antipolis France

*University of Massachusetts Amherst, USA

2 / 21

Video Streaming Services

Containers

Desktop Browsers Native Mobile Applications

What are the Network Characteristics of Video Streaming Traffic?

3 / 21

Objective

• What exactly happens during video streaming?– Arrival of data packets– Strategies to stream videos– Potential Impact

4 / 21

• Introduction and Motivation• Datasets and Measurement Techniques• Streaming Strategies• Impact of Streaming Strategies

Outline

5 / 21

Datasets

• YouTube videos– Flash, HTML5, and HD (Flash)– Mobile

• Netflix videos - Silverlight– Desktop – Mobile

6 / 21

Measurement Technique

802.11

Packet Capture

7 / 21

Measurement Locations

• France– Academic (Wired; Wi-fi for mobile)– Residential (Wi-fi)

• USA– Academic (Wired; Wi-fi for mobile)– Residential (Wired)

YouTube

YouTube and

Netflix

Similar Traffic Characteristics at Each Location

8 / 21

Outline• Introduction and Motivation• Datasets and Measurement Techniques• Streaming Strategies• Impact of Streaming Strategies

9 / 21

Generic Behavior of Video StreamingDo

wnl

oad

Amou

nt

Time

Buffe

ring

Block Size

On

OffSteady State

Average rate

∝Video encoding rate

10 / 21

We Identified Three Streaming Strategies

No On Off Cycles

Long On Off Cycles

OFF OFF

Short On Off Cycles

Streaming strategies vastly different

11 / 21

Streaming Strategies UsedService YouTube Netflix

Container Flash HD (Flash) HTML5 Silverlight

IE 9 Short No Short Short

Firefox Short No No Short

Chrome Short No Long Short

iOS (native)

- - Based on encoding rate

Short

Android (native)

- - Long Long

Streaming strategy differs withapplication type and container

12 / 21

Features Controlling Arrival of Data Packets

• Buffering Amount• Block Size• Accumulation Ratio

Average download rate in steady state phase

Video encoding rate=

13 / 21

Arrival of Packets for Short ON OFF Strategy

64 kB

40 sec. of playback

Server side rate control with

absence of ACK clocks

1.25Buffering

independent ofencoding rate

Browser throttles rate

256 kB

Significant differences between implementations

14 / 21

Outline• Introduction and Motivation• Datasets and Measurement Techniques• Streaming Strategies• Impact of Streaming Strategies

15 / 21

Impact of Streaming StrategiesNo On Off Long On Off Short On Off

TCP Friendly Yes – TCP File Transfer

Yes – Periodic File Transfer

Unknown traffic not ack-clocked

Playout buffer occupancy

Large Moderate Small

Unused bytes on user interruptions

Large amount

Moderate amount

Small amount

StrategyMetric

16 / 21

Model for Aggregate Rate of Streaming Traffic

• Objective– Capture statistical properties of aggregate

streaming trafficBarakat et al., A flow-based model for Internet backbone traffic, In IMW’02.

• Uses– Dimension the network– Quantify impact of user interruptions

17 / 21

Aggregate Rate of Video Streaming Traffic

Aggregate Rate

Arrival Rate of streaming sessions

(Poisson)

Amount of datadownloaded

18 / 21

Insights from Model

• No User Interruptions– Aggregate rate (mean, variance, etc.) independent

of streaming strategy– Dimensioning rules do not change– Strategy to optimize other goals (server load, etc.)

• Users Interruptions– Impact of buffering amount and accumulation

ratio on wasted bandwidth

19 / 21

Summary

• Most popular clients and containers for video streaming

• Streaming strategy differs with client applications and container– HTML5 streaming vastly differs with client

applications• Model to study impact of streaming

strategies

20 / 21

Open Questions for the CCN community

• Should CCN nodes be aware of the underlying streaming strategy?

• What is the optimal streaming strategy for CCN?

• Is there an optimal caching strategy for a given streaming strategy?

• What is the impact of user interruptions due to lack of interest on CCN caches?

21 / 21

THANK YOU

ashwin.rao@inria.fr

Network Characteristics of Video Streaming Traffic

B-22

BACKUPS

B-23

Short or Long• Block Size – Threshold 2.5 MB

Short Long

OFF

Long

B-24

ACK Clocks

• Source sends packets on receiving ACK• ACKs as an indication of available

bandwidth46 packets sent in the first RTT after an OFF period of more than 500 ms

25 / 21

Conclusion

• Most popular clients and containers for video streaming

• Streaming strategy differs with client applications and container– HTML5 streaming vastly differs with client

applications• Model to study impact of streaming

strategies

B-26

User Interruptions

Video duration

Playback time downloaded in buffering phase

1 Accumulation ratio Fraction of video watched

- X>

Video download will be in progress when

B-27

Impact of Losses

• Merging of cycles• Playback can freeze• Longer buffering phase

B-28

HTML5

• Primary - webM• Very few - h.264

B-29

Netflix Streaming StrategiesContainer Silverlight Silverlight for Mobile Devices

Application Any Web Browser

iOS (native) Android (native)

Strategy Short Short Long

Buffering Amount

30 MB to 150 MB

10 to 20 MB 35 to 45 MB

Block Size 0.5 MB to 2 MB

0.5 to 2.5 MB 4.5 to 6 MB

B-30

YouTube Streaming StrategiesContainer Flash HTML5

Application Any Web Browser

IE 9 Firefox Google Chrome

iOS (native)

Android (native)

Strategy Short Short No Long Multiple Long

Buffering Amount

40 s Up to 15 MB

Video Size

Up to 15 MB

40 s of playback or up to 20 MB

Up to 10 MB

Block Size 64 kB 256 kB NA 5 MB to 8 MB

64 kB 2 MB to 8 MB

B-31

Tradeoff

• Migration from one strategy to another can have a non-negligible impact

Raw File Transfer vs

Periodic Buffering vs

No ack-clock

B-32

Video Streaming in the Internet

• 20 % to 40 % of all Internet traffic– Traffic share steadily increasing in recent

years• Streaming over HTTP – using TCP

– Firewall configurations– TCP flows assumed to be fair